✦ Superintelligence Labs & the Talent BetThin moat
Meta Platforms (META) — the future bets
Fourteen billion for Scale's founder and nine-figure offers bought one shipped model so far — the frontier is rented until Avocado proves otherwise.
The bet began as an admission: Llama 4 landed with a shrug, rivals' open models passed it on benchmarks, and Zuckerberg responded the way he responds — with money. The $14.3 billion Scale AI investment brought Alexandr Wang to run the new Meta Superintelligence Labs, followed by a hiring campaign whose offers reached nine figures and a reorganization that shifted Meta away from its open-source identity1. In April 2026 the lab shipped its first model, Muse Spark — proprietary, with open-sourcing merely 'hoped for' in future versions — and next-generation models code-named Mango and Avocado are in the pipeline2.
The cost is already visible: research and development rose 67% to $21.7 billion in the June 2026 quarter, driven by compensation, data-centre and cloud costs, and third-party AI token costs.3 The wager is that talent density plus unmatched compute can buy back a frontier position that organic effort lost. The cost of being wrong is not just the billions but the ad engine's future: Meta's ranking, targeting and generative-ad roadmap all assume in-house models near the frontier. Watch two things: where Muse Spark and Avocado land on independent benchmarks against OpenAI, Google and Anthropic, and whether the famously expensive hires stay — talent that can command nine figures can command them somewhere else too.
A reset, honestly graded: $14.3B and a poaching spree produced a reorganization, a strategy shift away from open source, and one shipped model the leaderboards haven't crowned. Stable is what a rebuilding year looks like — the arrow moves when Avocado lands against the frontier.
Compensation, data centres, cloud services and third-party AI tokens drove the rise. R&D growing faster than revenue for long, without a model that pulls ahead, would mean the talent bet is buying cost rather than capability.
Source: Meta Form 10-Q, quarter ended 30 June 2026 ↗- ReportedThe $14.3 billion Scale AI investment brought Alexandr Wang to run the new Meta Superintelligence Labs, followed by a hiring campaign whose offers reached nine figures and a reorganization that shifted Meta away from its open-source identity.CNBC — a year after Meta's $14.3B Scale AI investment brought Alexandr Wang to lead Superintelligence Labs: the nine-figure hiring blitz, the reorganization, and the shift away from open source — Jun 2025 – Jun 2026 · publ. Jun 14, 2026 · source ↗
- ReportedIn April 2026 the lab shipped its first model, Muse Spark — proprietary, with open-sourcing merely 'hoped for' in future versions — and next-generation models code-named Mango and Avocado are in the pipeline.CNBC — Meta debuts Muse Spark (Apr 8, 2026): the first major model from Meta Superintelligence Labs since the $14.3B Scale AI deal brought in Alexandr Wang; proprietary, with open-sourcing only 'hoped for' in future versions; Mango (image/video) and Avocado (next-gen LLM) in the pipeline — Apr 2026 · publ. Apr 8, 2026 · source ↗
- ReportedThe cost is already visible: research and development rose 67% to $21.7 billion in the June 2026 quarter, driven by compensation, data-centre and cloud costs, and third-party AI token costs.Meta Form 10-Q, quarter ended 30 June 2026 - DAP 3.60B in June 2026 from 3.48B (+3%), the Q1 dip due to internet disruptions in Iran and restricted WhatsApp access in Russia; ARPP $16.86 (+24%); revenue by customer address US & Canada $23,863M, Europe $14,009M, Asia-Pacific $16,073M, Rest of World $6,856M; R&D $21,656M (+67%) including third-party AI token costs; FTC v. Meta: trial April-May 2025, judgment for Meta on 18 November 2025, FTC notice of appeal 20 January 2026; resellers serving China-based advertisers risk factor — Q2 2026 · publ. July 30, 2026 · source ↗
- Meta Platforms Form 10-K filings — Business & Risk Factors (SEC EDGAR)
- A year of Superintelligence Labs (CNBC)