At a glance: Meta Platforms. Restructuring announced and executed May 20, 2026. Scale: 8,000 employees laid off (10% of 78,865-person workforce); 6,000 open requisitions cancelled; ~7,000 employees transferred to AI-focused teams. Q1 2026 revenue: $56.31B. Net income: $26.8B. 2026 capex guidance: $125–145B. Part of our AI Models & Companies reviews.
On May 20, 2026, the same week Meta posted the highest quarterly revenue in its history, the company sent layoff notifications to 8,000 employees.
That tension — record profits, mass reductions — is not a contradiction. It is the point.
Zuckerberg has been telegraphing this trade since early 2025: use Meta’s profitable advertising engine to fund an AI infrastructure build that no other company can match at this pace, then staff the resulting AI-native company with fewer but more specialized people. May 20 is the execution of that plan.
The Numbers
Revenue and profit. Q1 2026 net income of $26.8 billion is the highest Meta has recorded in any quarter. Revenue of $56.31 billion represents continued growth off a strong 2025 base — full-year 2025 revenue was $201 billion, up 22% year-over-year, with free cash flow of $43.6 billion.
The layoffs. 8,000 employees (10% of Meta’s 78,865-person workforce) received notifications on May 20. An additional 6,000 open job requisitions were simultaneously cancelled, bringing the effective headcount reduction to 14,000 positions. Further cuts are planned for the second half of 2026.
Savings. Bank of America estimates the reductions generate $7–8 billion in annualized savings. That figure covers roughly 5–6% of the 2026 capex plan.
Capital expenditure. Meta raised its 2026 capex guidance to $125–145 billion — up from $72.2 billion in 2025 and $39.2 billion in 2024. In a single quarter, Meta added $107 billion in new contractual commitments for cloud and infrastructure deals. The capex plan is not being funded by the layoffs; it is being funded by the advertising revenue. The layoffs are about organizational composition, not financial necessity.
What the Layoffs Are Actually For
The framing in internal communications — confirmed by multiple reports — is not “we need to cut costs.” It is “we need a different kind of company.”
About 7,000 employees are being transferred, not cut, into three new AI-focused structures:
Applied AI Engineering (AAI) — engineers building AI agents that can autonomously carry out tasks currently performed by human staff. The explicit goal is internal automation: Meta agents doing Meta work.
Agent Transformation Accelerator (ATA XFN) — a cross-functional team coordinating how the rest of the company deploys AI agents into existing workflows. Where AAI builds the tools, ATA manages the rollout.
Central Analytics — measures productivity outcomes as agents replace human functions, creating the feedback loop for what works and what doesn’t.
The management structure is also flattening. Teams are reorganizing into smaller pods with higher autonomy. Managerial layers are being reduced. The target operating model resembles how software companies with AI-generated output can work: fewer coordinators, more direct contributors, agents handling the connective tissue.
The people being cut, by most accounts, are from roles that AI is expected to absorb in 12–24 months: certain data labeling operations, content moderation support, mid-layer program management, roles adjacent to the tasks being automated.
Prometheus: The Infrastructure Bet
The $145B capex is not abstract. The largest single piece is Prometheus — a 1-gigawatt AI supercluster under construction in New Albany, Ohio.
Prometheus will be the world’s first gigawatt-capable data center when it comes online in 2026. The facility is being built at 1500 Beech Road; Zuckerberg announced in July 2025 that it would be the centerpiece of Meta’s advanced AI infrastructure.
Power. Meta has signed nuclear energy deals with Vistra, TerraPower, and Oklo to power Prometheus. A 200-megawatt natural gas facility (from Will-Power OH) is also approved exclusively for the site, with a second 200 MW facility in the permitting process. Nuclear is the only power source capable of delivering baseload power at this density reliably enough for 24/7 AI training workloads.
Hardware. Prometheus will be GPU-dense from the ground up: custom AMD Instinct MI450-based chips, 6th Gen AMD EPYC CPUs, ROCm software, and AMD’s Helios rack-scale architecture. First shipments are scheduled for the second half of 2026.
Scale trajectory. Prometheus is not the ceiling. Meta has separately disclosed plans for multiple multi-gigawatt data center clusters, with a 5 GW total facility footprint on the horizon. Zuckerberg described one cluster as covering “a significant part of the footprint of Manhattan.”
Louisiana joint venture. Alongside Prometheus, Meta has committed to a $27 billion joint venture with Nebius for a gigawatt-scale data center campus in Louisiana. The JV structure distributes capital risk while keeping Meta in control of the resulting compute.
Alexandr Wang and Meta Superintelligence Labs
The organizational story behind the layoffs involves a leadership power structure that is more complicated than the org chart suggests.
Alexandr Wang — 28-year-old former CEO of Scale AI — joined Meta in June 2025 as Chief AI Officer. He runs Meta Superintelligence Labs (MSL), the division responsible for frontier model development. MSL’s first major external model, Muse Spark, was released earlier in May 2026.
MSL was reorganized in August 2025 into four groups: a research lab (TBD Lab), FAIR (Meta’s longtime AI research division), Products and Applied Research, and MSL Infrastructure. The AGI foundations team was dissolved into MSL at the same time.
In March 2026, a parallel structure was added: a new Applied AI Engineering unit under Maher Saba, reporting to CTO Andrew Bosworth rather than to Wang. The effect is a dual-track AI leadership model. Wang runs frontier model development and research; Saba runs internal AI deployment. Both report upward through different chains.
Multiple reports describe this as a hedge — Zuckerberg getting frontier AI ambitions (Wang) without making a single external hire the bottleneck for all AI deployment decisions.
The Paradox That Isn’t
The surface read on Meta May 2026 is uncomfortable: a company posting record profits cutting 10% of its workforce. The framing of “workers training their AI replacements” has appeared in several headlines.
The internal logic is consistent, even if the human cost is not comfortable. Meta’s advertising business is generating $43+ billion in annual free cash flow. That engine does not require 80,000 employees to operate at its current efficiency. What it can do is fund a compute infrastructure build that would be impossible for any company operating at a loss or near breakeven — which describes most of the competition in foundation models.
The $145B capex plan is the bet: that whoever has the most training compute by 2027–2028 wins the next round of model capability. Meta is funding that bet with advertising profit. The layoffs are not the source of funding. They are the organizational consequence of a company that has decided its future workforce looks different than its current one.
Whether the bet pays off depends on whether raw compute advantage still compounds into model quality at the frontier. The trend through 2025 — diminishing returns on scale-only approaches — leaves that open. Meta’s counterargument, embodied in Prometheus, is that no one has yet trained at gigawatt scale. The outcome at that scale is genuinely unknown.
What to Watch
- H2 2026 cuts — Additional layoffs are confirmed for the second half of the year. The scope has not been announced.
- Prometheus go-live — First shipments in H2 2026; full operational status is the real test of whether 1 GW training translates to capability gains.
- Muse Spark benchmarks — MSL’s first major public model; performance data against Claude, GPT, and Gemini at the frontier will be the first signal of whether Wang’s team is producing at the level Zuckerberg hired him for.
- Agent deployment metrics — The ATA team is meant to measure agent productivity replacing human work. If Meta publishes this data (or if it leaks), it will be the most concrete evidence yet of whether internal AI agents at scale are delivering.
- Wang vs. Bosworth — The dual-track AI leadership structure is inherently unstable over a 12–18 month horizon. One track will likely absorb the other. Which one determines Meta’s AI identity for the next phase.