Meta Cut 600 From FAIR and MSL Infra. TBD Lab Kept the Money.

On October 22, 2025, Meta cut roughly 600 roles from Meta Superintelligence Labs (MSL), concentrated in FAIR, Products & Applied Research, and MSL Infra. Alexandr Wang's Artificial Superintelligence Lab (TBD Lab) was explicitly excluded. Four weeks later, Yann LeCun confirmed his departure to launch AMI Labs in Paris. This is the cleanest, most legible ex-hyperscaler AI talent pool of the cycle, and the sourcing window has a hard end date.

The 600-person cut, mapped to sub-org

Meta's cut hit three of the four MSL sub-orgs formed in the August 2025 reorg: FAIR, Products & Applied Research, and MSL Infra. TBD Lab, the shielded Wang-led team, was untouched. Post-cut MSL headcount sits just under 3,000, with terminations dated Nov 21, 2025 and a non-working notice period that removes internal access on day one.

The reorg matters because it tells you which badge to search for. Anyone still listing "Meta AI" or "GenAI" on their profile is behind. The four sub-orgs are:

• FAIR (Fundamental AI Research), historically around 1,000 headcount, research-first, unwinding since Joelle Pineau's exit on May 30, 2025.
• Artificial Superintelligence Lab / TBD Lab, Wang's shielded high-comp team, backed by the $14.3B Scale AI investment and co-led by Nat Friedman.
• Products & Applied Research, the applied teams shipping into Instagram, WhatsApp, and Meta AI.
• MSL Infra, the training and inference platform.

Three of those four are the sourcing pool. The fourth is the compensation gravity you cannot beat.

~600

Why the severance clock ends in mid-March 2026

The sourcing window closes around mid-March 2026, because severance is 16 weeks plus 2 weeks per year of service, and the termination date is Nov 21, 2025. That is the "cheap money" window before candidates re-anchor to top-of-market comp at AMI, Thinking Machines Lab, or Anthropic.

A departing MSL engineer with 4 years of service is sitting on 24 weeks of pay. That runway ends in early May 2026. The behavioral shift happens earlier, around week 12 to 14, when candidates start actively negotiating instead of listening. AMI Labs closed its $1.03B seed at a $3.5B pre-money on March 10, 2026, the largest European seed on record. That is not a coincidence in timing. LeCun timed his fundraise to the end of the Meta severance runway.

If you are doing Meta FAIR layoffs sourcing in Q2 2026, you are picking through what AMI, Mistral, and Thinking Machines already rejected. Every week you wait past January costs you a percentile.

What the ex-Meta AI cohort actually looks like in-market

The ex-Meta AI researcher slice is small, legible, and already mid-migration. A filtered pull on the "ex-FAIR" keyword plus PyTorch and deep learning skills returns 10 named researchers, and 40% of them are already at direct rivals or LeCun's new lab.

These are sample-level counts from a filtered slice, not a full market census. What they tell you: the destinations are stable (DeepMind, OpenAI, Databricks, AMI), the geographies are concentrated (Bay Area, Paris, NYC, London), and Paris is punching above its weight because LeCun opened FAIR-Paris in 2015 and personally recruited most of it.

Describing that in a Boolean search takes 30 minutes and misses half the pool. Describing it in one English sentence is what Refolk does: "ex-Meta FAIR research engineers in Paris or NYC with production LLM training experience, not currently at TBD Lab" returns a ranked list, and you spend the saved 25 minutes on the first messages.

MSL Infra is the under-sourced half of this

Every recruiter is chasing ex-FAIR researchers. The scarcer profile is the ex-MSL Infra engineer who ran Llama-scale training clusters, and almost none of them self-label as "MSL Infra" on public profiles. This is the arbitrage.

A search on the literal "MSL Infra" title string returns zero hits. The profile exists, it just is not tagged that way. To find these engineers, you have to source on surrogate signals:

• Systems keywords: NCCL, RDMA, InfiniBand, Slurm at scale, PyTorch FSDP, DeepSpeed, Megatron, vLLM.
• Hardware rollout evidence: GB200, H200, TPU v5, custom accelerator bring-up.
• Company + tenure: Meta 2022 to 2025, with GitHub activity on PyTorch, torchtitan, torchtune, or fairscale.
• Talk / paper trail: MLSys, OSDI, or internal Meta engineering blog authorship.

The scarcity is org-chart optics. Infra engineers at hyperscalers routinely describe themselves as "Software Engineer" or "Production Engineer" for compensation-band reasons. Title strings will not find them. Skill graphs and repo activity will.

Yann LeCun and AMI Labs: the downstream vacuum

Treat LeCun's AMI Labs as a leading indicator, not a competitor. AMI is HQ Paris with offices in NYC, Montreal, and Singapore, matching FAIR's historical footprint almost exactly, and LeCun personally recruited most of FAIR-Paris a decade ago. Expect a meaningful share of FAIR-Paris research staff to test AMI within 12 months; the Refolk sample already shows 1 of 10 FAIR-tagged researchers at AMI within four months of its announcement.

The AMI cap table (Cathay Innovation, Greycroft, Hiro Capital, HV Capital, Bezos Expeditions, plus NVIDIA, Samsung, and Toyota Ventures as strategics) buys world-models research at a $3.5B pre-money valuation. That is a comp anchor. Any ex-Meta AI recruiting conversation you run in Q1 2026 has to answer "how does this compare to AMI?" whether or not the candidate has an AMI offer, because their peers are quoting it in Slack.

The tactical takeaway: use LeCun's personal network, and Alexandre LeBrun's (AMI's CEO, previously founder of Nabla), as a source-of-sources map. Named FAIR alumni who have already moved make warm-intro targets:

• Naman Goyal and Shaojie Bai, to Thinking Machines Lab (March 2025).
• Dan Bikel, to Writer as Head of AI (June 2025).
• Chi-Hao Wu, to Memories.ai as CAIO.
• Cristian Canton, to Barcelona Supercomputing Center.
• Joelle Pineau (ex-FAIR VP), whose 8-year network is the most valuable warm-intro map in the cohort.

The people you cannot poach from TBD Lab are the wrong anchor. The 600 you can are the correct one.

The founder-track read

If you are a founder rather than a recruiter, the 2025 to 2026 cut cohort is your cofounder pool, not your first-hire pool. Every prior FAIR unwind produced a generation of AI companies: Perplexity, Mistral, Fireworks AI, and World Labs all trace back to earlier FAIR or Meta GenAI departures.

The pattern repeats because FAIR selects for a specific profile: senior researcher with production-scale systems context and enough autonomy to have owned a paper end-to-end. That is also the cofounder profile. If you are seed-stage and looking for a technical cofounder in Meta Superintelligence Labs alumni, this is the highest-density window in years. Move in November and December 2025, not March 2026.

A concrete sourcing playbook for the next 90 days

Run the same play in the same order for every named ex-MSL candidate. The reason it works is severance economics: every candidate is on the same clock, anchored on Nov 21, 2025.

1. Week 1 to 2: pull the named list. Filter by sub-org (FAIR, Products, MSL Infra), geography (Bay Area, Paris, NYC, London), and destination-not-yet-declared (LinkedIn still says Meta, or "open to work" with no new logo).
2. Week 3 to 4: first-touch messaging. Reference the sub-org they were in, not "Meta AI" generically. Ex-FAIR responds to research-mission language; MSL Infra responds to infrastructure-scale language.
3. Week 5 to 8: intro calls, no pitch. Use them to map the second-degree network. Every candidate can name 3 to 5 peers still deciding.
4. Week 9 to 12: convert the peer names into a warm-intro queue. The FAIR/MSL network is small and highly interconnected, so this compounds fast.
5. Week 13 to 16: close on the ones aligned to your comp band. The rest go into a tracked list for a re-touch at week 20, when severance runs out and cash pressure resets expectations.

Skip step 3, and you get 30% of the value. Every senior candidate is worth two more names, and those names are cheaper to convert than a cold pull.

FAQ

How do I tell if an ex-Meta AI candidate was in TBD Lab or in the cut sub-orgs?

TBD Lab is small, Wang-led, and external-hire heavy, and its members did not change titles in October 2025. The cut cohort will show a Nov 21, 2025 end date at Meta and, in most cases, a public post or "open to work" flag within two weeks of that date. If a candidate is still at Meta in December 2025 in an AI role, assume TBD or an internal redeployment, and treat the comp anchor accordingly. If they left on Nov 21 with no new logo yet, they are in the sourcing window.

Is FAIR still functioning after the cut?

FAIR still exists on the org chart, but its research custody has moved into TBD Lab. Post-Pineau (May 2025), post-LeCun (November 2025), and post-October cut, the remaining FAIR structure is smaller, applied, and no longer the center of gravity for Meta's frontier work. That is why the alumni network is the interesting artifact, not the current org.

What comp should I benchmark against for ex-MSL candidates?

Do not benchmark against TBD Lab. Those packages are a shield structure, not a market. Benchmark against Thinking Machines Lab, Anthropic, and AMI Labs at $3.5B pre-money. For MSL Infra engineers hiring in Q1 2026 specifically, the closer comp is senior infra at OpenAI or DeepMind, plus a startup equity premium if you are pre-Series-B.

Can I build this list without knowing the sub-org names?

Yes. Describe the person in plain English ("ex-Meta AI researchers laid off in the October 2025 cut, based in Paris or Bay Area, with LLM training experience") and Refolk returns a ranked shortlist across GitHub, LinkedIn, and the open web. The sub-org filtering, surrogate keywords, and geographic weighting are handled by the query, not by you.